Multimodal fusion and vision–language models: A survey for robot vision

In the paper

“DefFusion [14] introduces a deformable transformer mechanism that dynamically adjusts to multimodal data, mitigating over-attention and noise, and significantly improving 3D semantic segmentation, particularly on large-scale benchmarks like NuScenes.”

What the sources say

• Disputed

  Section III describes the proposed method as Multi-Sem Fusion, using adaptive-attention and deep-feature-fusion modules, not a deformable transformer called DefFusion. The contributions passage and Table III report improvements in 3D object detection on nuScenes; semantic segmentation supplies inputs rather than being the reported outcome. Our copy does not substantiate the claimed mitigation of over-attention and noise. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [14] Xu, 2024, DefFusion: Deformable multimodal representation fusion for 3D semantic segmentation sourceExperimental — accuracy not yet measured

In the paper

“In addition, although Wu et al. [31] provide an in-depth analysis of multimodal techniques for embodied navigation, their work does not include emerging vision-language model architectures, nor does it place tasks such as SLAM and manipulation within a unified analytical framework.”

What the sources say

• Disputed

  The cited paper is Bordes et al.’s introduction to vision-language models, not the Wu et al. work described in the claim. Its Introduction explicitly says it presents VLM training paradigms and models using pretrained backbones, contrary to the claim that the cited work omits VLM architectures. Our copy does not establish what Wu et al. cover. (LLM, conf=0.84)

  Checked by a reasoning model — may be wrong.

  cites [31] Bordes, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Lastly, Multi-Sem Fusion [60] integrates LiDAR and camera data for 3D object detection, employing adaptive attention and deep feature fusion to enhance detection accuracy, although its reliance on offline processing limits real-time applicability.”

What the sources say

• Disputed

  The paper’s opening summary (p. 1) describes DDRNets for real-time semantic segmentation, not Multi-Sem Fusion for LiDAR–camera 3D object detection, and reports 102 FPS on Cityscapes. The Introduction (p. 2) says the method achieves its results without an attention mechanism. Its real-time performance directly contradicts the claimed reliance on offline processing. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  cites [60] Hong, 2021 sourceExperimental — accuracy not yet measured

In the paper

“MuSe-GNN [67] further integrates modality-specific entities into a unified semantic space by leveraging shared GNN encoders and jointly optimizing similarity-and contrastive-based objectives.”

What the sources say

• Disputed

  The conclusion (p. 11) identifies this paper’s contribution as the squeeze-and-excitation (SE) block for channel-wise feature recalibration, not MuSe-GNN. The introduction (p. 1) likewise describes a convolutional-network architecture rather than shared GNN encoders or similarity- and contrastive-based objectives. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  cites [67] J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141. sourceExperimental — accuracy not yet measured

In the paper

“CenterFusion [94] uses a staccato-based method to correlate radar data with camera cen-troids, enhancing 3D object detection accuracy, especially in dynamic environments by integrating velocity and depth information.”

What the sources say

• Disputed

  The Introduction and Section 3 identify the paper’s method as RRPN, which uses radar detections to generate regions of interest—not CenterFusion or a staccato-based method correlating radar data with camera centroids. Section 3 does say radar range and range-rate can supply depth and velocity information, but it does not substantiate the claimed CenterFusion method or its 3D detection result. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [94] Nabati, 2019, Rrpn: Radar region proposal network for object detection in autonomous vehicles sourceExperimental — accuracy not yet measured

In the paper

“The CRAFT [100] model adapts radar-camera correlation in polar coordinates, utilizing cross-attention for robust fusion.”

What the sources say

• Disputed

  Section 4 (p. 4) identifies the paper’s model as CenterFusion and describes associating radar detections with object centers for middle fusion. The Introduction (p. 2) specifies frustum-based association and radar feature maps projected onto the image plane, rather than the claim’s CRAFT model with polar-coordinate correlation and cross-attention. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [100] R. Nabati, H. Qi, Centerfusion: Center-based radar and camera fusion for 3d object detection, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 1527–1536. sourceExperimental — accuracy not yet measured

In the paper

“Wang et al. [149] adopted hybrid neural representations and multi-resolution hash grids to achieve high-fidelity scene reconstruction, significantly improving accuracy and robustness in RGB-D SLAM tasks.”

What the sources say

• Disputed

  The Abstract (page not provided) identifies this as Shan et al.’s LVI-SAM, a factor-graph system coupling lidar, visual, and inertial odometry—not a Wang et al. method using hybrid neural representations or multi-resolution hash grids. The Experiments section describes tests of LVI-SAM, not high-fidelity reconstruction for RGB-D SLAM. (LLM, conf=0.89)

  Checked by a reasoning model — may be wrong.

  cites [149] Shan, 2021, Lvi-sam: Tightly-coupled lidar-visual-inertial odometry via smoothing and mapping sourceExperimental — accuracy not yet measured

In the paper

“The 3D-VLA [154] model enhances visual representation by generating 3D point clouds and depth maps, improving task precision and robustness with language-guided action generation in dynamic environments.”

What the sources say

• Disputed

  The map-representation section (p. 4) identifies the system as ConceptFusion and says it takes color and depth images as input to build a 3D map; it does not describe generating depth maps. Related Work (p. 3) explicitly distinguishes its perception system from approaches focused on planning and control, contrary to the claim that this model generates language-guided actions. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [154] Jatavallabhula, 2023 sourceExperimental — accuracy not yet measured

In the paper

“BLIP [190], through a generative model, jointly trains images and text, enabling mutual generation and understanding between the two modalities, thus enhancing the performance of vision-language generation tasks.”

What the sources say

• Disputed

  The Introduction describes the paper's method as self-supervised learning from audio and video, and Figure 1 depicts clustering audio and visual features to train encoders. That is contrary to identifying this paper as BLIP or describing its method as generative image–text training for vision-language generation. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [190] Alwassel, 2020, Self-supervised learning by cross-modal audio-video clustering, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“ViT [198] introduced patch embedding, dividing images into small patches and converting them into tokens for input into the Transformer, enabling effective feature extraction through self-attention.”

What the sources say

• Disputed

  The Introduction (p. 2) says this paper introduces the Transformer for sequence transduction, not ViT or image patch embedding. The Conclusion (p. 10) describes applying the Transformer to images as future work, contrary to attributing an image-patch method to this paper. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [198] Vaswani, 2017, Attention is all you need, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“During training,the model learns the joint distribution between the two modalities and optimizes the model through methods such as adversarial training or maximizing likelihood, enabling it to generate the target modality”

What the sources say

• Disputed

  Section 2.3 (p. 4) says CLIP learns to identify which images and texts are paired, using a contrastive objective and symmetric cross-entropy loss. It explicitly distinguishes this from predicting the accompanying text; it does not describe adversarial training, maximizing likelihood to learn a joint distribution, or generating a target modality. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  cites [188] Radford, 2021, Learning transferable visual models from natural language supervision sourceExperimental — accuracy not yet measured

In the paper

“In contrast, multimodal fusion integrates complementary information from various sensor inputs (e.g., vision, language, depth, LiDAR, and tactile data), thereby enhancing the perception, reasoning, and decision-making capabilities of robot vision systems [4,5,6,7].”

What the sources say

• ?Not checked — no full text of the sourcecites [4] Xu, 2024, MRFTrans: Multimodal representation fusion transformer for monocular 3D semantic scene completion, Inf. Fusion source
• ?Not checked — no full text of the sourcecites [5] X. Chen, H. Ma, J. Wan, B. Li, T. Xia, Multi-view 3d object detection network for autonomous driving, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 1907–1915. source
• Partly supported

  The Introduction (p. 1) says fusing visual and inertial information improves robustness in complex environments, and §2.1 (p. 2) says that fusion helps recover metric scale. The paper does not support the broader examples of language, depth, LiDAR, or tactile inputs, or the claimed improvements in reasoning and decision-making. (LLM, conf=0.88)

  cites [6] Li, 2024, RD-VIO: Robust visual-inertial odometry for mobile augmented reality in dynamic environments, IEEE Trans. Vis. Comput. Graphics sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [7] Babadian, 2023, Fusion of tactile and visual information in deep learning models for object recognition, Inf. Fusion source

In the paper

“Large-scale pretrained VLMs not only possess cross-modal alignment and generalization abilities, but also demonstrate strong potential in tasks such as zero-shot understanding [8], instruction following [9,10], and visual question answering [11].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [8] Radford, 2019, Language models are unsupervised multitask learners, OpenAI Blog
• Partly supported

  Section 2.2, “Large Models in Robotics Navigation” (page not provided), describes multimodal large models as capable of instruction following and visual perception, with strong generalization. Section 4.3 reports zero-shot visual-language navigation. The paper does not establish the broader claim of zero-shot understanding or explicitly discuss cross-modal alignment; visual question answering is attached to a separate citation in the citing sentence. (LLM, conf=0.82)

  cites [9] Long, 2024 sourceExperimental — accuracy not yet measured
• ?Could not verify

  The citation is attached to the claim about VLMs’ zero-shot understanding. The Introduction (p. 3) discusses zero-shot vision-and-language navigation using LLMs, but does not establish strong zero-shot understanding by pretrained VLMs in general. A search for zero-shot VLM understanding found no passage addressing that claim. (LLM, conf=0.70)

  cites [10] Zhou, 2024, Navgpt-2: Unleashing navigational reasoning capability for large vision-language models sourceExperimental — accuracy not yet measured
• Supported

  Citation [11] is attached to visual question answering. Section 5.2.1 (Image Understanding) evaluates Gemini on question-answering tasks including VQAv2 and describes its zero-shot image-understanding performance, supporting the attributed point. (LLM, conf=0.93)

  cites [11] Team, 2023 source

In the paper

“Roitberg et al. [47] compared and analyzed seven decision-level fusion strategies for driver behavior understanding, offering valuable insights for selecting appropriate fusion schemes.”

What the sources say

• ?Could not verify

  The Multimodal Fusion Ablation Studies (p. 34) and Fusion Stage and Reliance on Modalities (p. 35) compare configurations of this paper’s semantic-segmentation model, not seven decision-level strategies for driver behavior understanding. Searches for Roitberg and the seven-strategy claim found no passage in our copy addressing it. (LLM, conf=0.75)

  cites [47] Valada, 2018, Self-supervised model adaptation for multimodal semantic segmentation, Int. J. Comput. Vis. sourceExperimental — accuracy not yet measured

In the paper

“As illustrated in Figure 3, multimodal fusion strategies can be categorized into early fusion, mid fusion, and late fusion based on the fusion strategy [38,39].”

What the sources say

• ?Not checked — no full text of the sourcecites [38] Zhang, 2020, Deep multimodal fusion for semantic image segmentation: A survey, Image Vis. Comput. source
• Supported

  Section 4 explicitly categorizes fusion strategies as early, feature-level, and late fusion according to whether fusion occurs at the input, an intermediate stage, or the end of processing. The claim's “mid fusion” corresponds to the paper's intermediate, feature-level category. (LLM, conf=0.94)

  cites [39] Rizzoli, 2022, Multimodal semantic segmentation in autonomous driving: A review of current approaches and future perspectives, Technologies source

In the paper

“On the other hand ,DMA [59] focuses on open-vocabulary 3D scene understanding, aligning dense correspondences between 3D points, 2D images, and text, utilizing vision-language models to generalize across diverse indoor and outdoor scenarios.”

What the sources say

• ?Could not verify

  The Semantic Segmentation section (p. 7) describes HRNetV2 assigning class labels to image pixels, not DMA. Searches for DMA, open-vocabulary 3D scene understanding, 3D–2D–text correspondences, and vision-language models found no passage addressing the attributed claim in our copy. (LLM, conf=0.77)

  cites [59] Wang, 2021, Deep high-resolution representation learning for visual recognition, IEEE Trans. Pattern Anal. Mach. Intell. sourceExperimental — accuracy not yet measured

In the paper

“In addition, adversarial representation learning has been used to create modality invariant embedding spaces, reduce modal gaps, and improve cross modal fusion.”

What the sources say

• Supported

  The Abstract (p. 1) says the authors use adversarial training to learn a modality-invariant embedding space and reduce the modality gap for cross-modal fusion. Joint Embedding Space Learning (p. 3) adds that narrowing the gap allows representations from different modalities to be fused directly. (LLM, conf=0.94)

  cites [44] Mai, 2019 source

In the paper

“It combines various sources of information, such as vision, language, depth data, lidar, and speech, to enhance the capability of scene semantic understanding [35,36,37].”

What the sources say

• ?Not checked — no full text of the sourcecites [35] Mikolajczyk, 2005, A performance evaluation of local descriptors, IEEE Trans. Pattern Anal. Mach. Intell. source
• Partly supported

  Section 5.1 describes fusion of multiple visual representations for image classification, and Section 6.1.1 says high-performing methods combined visual feature channels. The paper does not support the claimed combination of vision with language, depth, LiDAR, and speech for scene semantic understanding. Page numbers were not available in our copy. (LLM, conf=0.84)

  cites [36] Everingham, 2010, The pascal visual object classes (VOC) challenge, Int. J. Comput. Vis. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [37] Geiger, 2012, Are we ready for autonomous driving? The KITTI vision benchmark suite source

In the paper

“3D-CVF [81] employs adaptive gating fusion with attention maps for feature weighting.”

What the sources say

• ?Could not verify

  The Related Work section (p. 3) cites 3D-CVF as an example of instance-level fusion, and the References (p. 12) identify that work. Neither passage describes adaptive gating or attention-map feature weighting; searches for those terms found only descriptions of this paper’s own fusion and transformer attention. Our copy therefore does not verify the attributed detail about 3D-CVF. (LLM, conf=0.77)

  cites [81] Li, 2022, Unifying voxel-based representation with transformer for 3d object detection, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“Point-Painting [80] projects LiDAR point clouds onto the output of the image semantic segmentation network and concatenates the channel activation values with the LiDAR point cloud intensity values, effectively utilizing image semantic information.”

What the sources say

• ?Could not verify

  The Related Work section (p. 2) discusses point- and voxel-based detectors, but does not describe Point-Painting. Searches for PointPainting and for LiDAR–image semantic fusion found no passage explaining the claimed projection or concatenation of values. (LLM, conf=0.76)

  cites [80] J. Mao, Y. Xue, M. Niu, H. Bai, J. Feng, X. Liang, H. Xu, C. Xu, Voxel transformer for 3d object detection, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3164–3173. sourceExperimental — accuracy not yet measured

In the paper

“To enhance temporal consistency and modality alignment, CCA-GNN [68] proposes a temporal positional encoding scheme for graph nodes and aligns audio-visual representations through canonical correlation analysis.”

What the sources say

• ?Could not verify

  The Implementation Details section describes OccFusion’s camera, lidar, and radar framework, not CCA-GNN. Searches for CCA-GNN, temporal positional encoding, and audio-visual alignment by canonical correlation analysis found no passage addressing the claim. The References section identifies [68] as VoxFormer, but that does not establish what CCA-GNN proposes. (LLM, conf=0.78)

  cites [68] Ming, 2024, OccFusion: Multi-sensor fusion framework for 3D semantic occupancy prediction, IEEE Trans. Intell. Veh. sourceExperimental — accuracy not yet measured

In the paper

“In contrast, LiDAR-based detection [73,74,75,76,77] processes 3D point cloud data, offering precise depth estimation and spatial structure representation.”

What the sources say

• ?Could not verify

  Section 2.1 discusses representations of program code, and Section 4.2.2 discusses mapping code to CPUs and GPUs. Searches for LiDAR, 3D point clouds, depth sensing, and object detection found no passage addressing the attributed claim. The retrieved passages have no page numbers. (LLM, conf=0.78)

  cites [73] A. Dutta, J. Alcaraz, A. TehraniJamsaz, E. Cesar, A. Sikora, A. Jannesari, Performance optimization using multimodal modeling and heterogeneous GNN, in: Proceedings of the 32nd International Symposium on High-Performance Parallel and Distributed Computing, 2023, pp. 45–57. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [74] X. Yang, S. Feng, Y. Zhang, D. Wang, Multimodal sentiment detection based on multi-channel graph neural networks, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021, pp. 328–339. source
• Partly supported

  The abstract (p. 1) says AVOD processes LiDAR point clouds alongside RGB images to predict objects’ extents and orientation in 3D space. The feature-map method (p. 3) describes encoding point heights and density, supporting spatial-structure representation. The paper reports accurate 3D localization, but does not establish the claim of precise depth estimation by LiDAR-based detection specifically. (LLM, conf=0.84)

  cites [75] Ku, 2018, Joint 3d proposal generation and object detection from view aggregation sourceExperimental — accuracy not yet measured
• Partly supported

  Section 4 (page unavailable) says that a LiDAR-based 3D detector processes a point cloud and that ground-truth depth measurements come from LiDAR point clouds. It also describes point-cloud data as containing 3D coordinates, supporting the spatial-representation point. The copy does not establish the claimed precision of LiDAR-based depth estimation; its statement about high precision refers to a monocular depth network. (LLM, conf=0.78)

  cites [76] D. Park, R. Ambrus, V. Guizilini, J. Li, A. Gaidon, Is pseudo-lidar needed for monocular 3d object detection?, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3142–3152. sourceExperimental — accuracy not yet measured
• Supported

  The Introduction (p. 1) describes 3D object detection based on point clouds and states that LiDAR point clouds provide spatial and structural information, including precise depth and relative location. This supports the cited portion of the sentence. (LLM, conf=0.94)

  cites [77] Yang, 2018 source

In the paper

“DeepFusion [85] ensures precise alignment of LiDAR and image features using InverseAug and cross-attention mechanisms.”

What the sources say

• ?Could not verify

  The Introduction (p. 1) includes DeepFusion among methods that merge image and LiDAR features, and the References (p. 12) identify it as a separate paper. Our copy does not say that DeepFusion uses InverseAug or cross-attention to ensure precise alignment; searches for those terms found no supporting passage. (LLM, conf=0.76)

  cites [85] Yang, 2022, Deepinteraction: 3d object detection via modality interaction, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“Camera-based single-modal 3D object detection [1, 72, 2, 3] infers 3D information from 2D images.While RGB images provide rich visual data, the lack of direct depth perception requires complex estimation techniques, making detection vulnerable to occlusions, poor angles, and adverse weather conditions.”

What the sources say

• Partly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [1] Z. Xu, W. Zhang, X. Ye, X. Tan, W. Yang, S. Wen, E. Ding, A. Meng, L. Huang, Zoomnet: Part-aware adaptive zooming neural network for 3d object detection, in: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, 2020, pp. 12557–12564, 07. sourceExperimental — accuracy not yet measured
• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [72] Passos, 2023, Multimodal audio-visual information fusion using canonical-correlated graph neural network for energy-efficient speech enhancement, Inf. Fusion sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [2] X. Shi, Q. Ye, X. Chen, C. Chen, Z. Chen, T.-K. Kim, Geometry-based distance decomposition for monocular 3d object detection, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15172–15181. source
• ?Not checked — no full text of the sourcecites [3] L. Wang, L. Du, X. Ye, Y. Fu, G. Guo, X. Xue, J. Feng, L. Zhang, Depth-conditioned dynamic message propagation for monocular 3d object detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 454–463. source

In the paper

“Fusion-Net [92] uses dense 2D radar maps and spatial transformation to align sensor feature maps.”

What the sources say

• ?Could not verify

  The Multi-modal Fusion and Adaptive Attention Module sections describe FusionPainting’s LiDAR–camera fusion, not Fusion-Net. Searches for “Fusion-Net,” “dense 2D radar maps,” and “spatial transformation” found no passage addressing the attributed method; the available passages therefore neither support nor contradict the claim about Fusion-Net. (LLM, conf=0.76)

  cites [92] Xu, 2021, Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection sourceExperimental — accuracy not yet measured

In the paper

“However, its camera features lack spatial awareness, which GAFusion [89] improves using Sparse Depth Guidance and Laser Occupancy Guidance.”

What the sources say

• Partly supported

  Section C, “Fully-Convolutional Fusion” (page unavailable), says BEVFusion’s camera and LiDAR features can be spatially misaligned because of inaccurate depth, supporting a narrower version of the spatial-awareness concern. The paper describes its own convolution-based BEV encoder as a response; our copy does not identify GAFusion or describe Sparse Depth Guidance or Laser Occupancy Guidance. (LLM, conf=0.85)

  cites [89] Liu, 2023, Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation sourceExperimental — accuracy not yet measured

In the paper

“Simple-BEV [99] combines camera and radar features through rasterised radar information and bilinear sampling for 3D bird's-eye view representation.”

What the sources say

• ?Could not verify

  The Camera-Radar Fusion section discusses other fusion methods but does not mention Simple-BEV. Searches for Simple-BEV, rasterised radar information, and bilinear sampling found no passage in our copy that addresses the claimed method. (LLM, conf=0.75)

  cites [99] Y. Kim, S. Kim, J.W. Choi, D. Kum, Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer, in: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023, pp. 1160–1168, 1. sourceExperimental — accuracy not yet measured

In the paper

“Instructionfollowing Navigation relies on the development of Room-to-Room (R2R) and REVERIE tasks, which enhances natural language interaction and environment-specific reasoning, and further highlights the role of multimodal perception in complex scenarios [111,31,112,113].”

What the sources say

• ?Not checked — no full text of the sourcecites [111] M. Savva, A. Kadian, O. Maksymets, Y. Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V. Koltun, J. Malik, et al., Habitat: A platform for embodied ai research, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 9339–9347. source
• ?Could not verify

  The passages in §3.5 discuss instruction tuning and multimodal instruction following, but not instruction-following navigation. Searches for “R2R,” “Room-to-Room,” and “REVERIE” found no passage in our copy connecting those tasks to natural-language interaction, environment-specific reasoning, or multimodal perception in navigation. (LLM, conf=0.76)

  cites [31] Bordes, 2024 sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [112] Zhu, 2017, Target-driven visual navigation in indoor scenes using deep reinforcement learning source
• Partly supported

  Section 4.1 (page unavailable) describes navigation to an object given language instructions, and the Introduction discusses environmental perception and task reasoning. But our copy does not identify R2R or REVERIE as tasks underlying instruction-following navigation or attribute the claimed advances to them; searches for both names found no relevant passage. (LLM, conf=0.86)

  cites [113] Ren, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Meanwhile, ViNG [17] enables visual goal navigation by constructing a topological graph, allowing robots to operate without pre-built maps or explicit spatial priors.”

What the sources say

• Supported

  The Introduction (pp. 1–2) says ViNG plans over a learned topological graph of past observations and navigates to visual goals without access to spatial maps. It also says the method makes no geometric assumptions about the environment and does not assume ground-truth spatial information. (LLM, conf=0.94)

  cites [17] Shah, 2021, Ving: Learning open-world navigation with visual goals source

In the paper

“UVTR [84] converts images into voxel representations, enabling cross-modal learning.”

What the sources say

• Partly supported

  The Introduction and C. Multimodal Fusion (page not provided) describe MVX-Net combining image features with LiDAR features at 3D points or voxels, supporting the cross-modal learning aspect. They do not describe UVTR or say that images are converted into voxel representations; VoxelFusion instead appends image features to features of voxels formed from point clouds. (LLM, conf=0.86)

  cites [84] Sindagi, 2019, Mvx-net: Multimodal voxelnet for 3d object detection sourceExperimental — accuracy not yet measured

In the paper

“Additionally, it employs a Laser-Guided Adaptive Fusion Transformer and a temporal fusion module for multi-frame aggregation.”

What the sources say

• ?Could not verify

  The Method overview and Fully-Convolutional Fusion section describe BEVFusion’s shared BEV representation and convolution-based encoder, not GAFusion. Searches for GAFusion, the Laser-Guided Adaptive Fusion Transformer, and a temporal fusion module found no passage in our copy addressing those components or their use for multi-frame aggregation. (LLM, conf=0.76)

  cites [89] Liu, 2023, Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation sourceExperimental — accuracy not yet measured

In the paper

“For example, the RT-2 [152] model aligns visual and language representations via pretraining, allowing the robot to generate control commands from linguistic input using a Transformer architecture to map high-level semantics to actions.”

What the sources say

• ?Could not verify

  The Introduction and Conclusion describe UVIO as a UWB-aided visual-inertial localization framework, not an RT-2 model. Searches for RT-2, vision-language pretraining, and Transformer-based language-to-action control found no passage addressing the claim in our copy. (LLM, conf=0.78)

  cites [152] Delama, 2023, UVIO: An UWB-aided visual-inertial odometry framework with bias-compensated anchors initialization sourceExperimental — accuracy not yet measured

In the paper

“Furthermore, LVI-SAM [150] represents a landmark multimodal SLAM framework, utilizing factor graph optimization to deeply couple visual and LiDAR data, particularly excelling in low-texture and dynamic environments.”

What the sources say

• ?Could not verify

  The Introduction and Related Work describes LIC-Fusion as a filtering approach and discusses graph optimization in other methods; it does not identify or describe LVI-SAM. The Vicon Room Sequences section reports LIC-Fusion 2.0 results in cluttered environments, not LVI-SAM performance in low-texture or dynamic environments. Searches for LVI-SAM, factor-graph coupling, and those environments found no passage addressing the attributed claim. (LLM, conf=0.76)

  cites [150] Zuo, 2020, Lic-fusion 2.0: Lidar-inertial-camera odometry with sliding-window plane-feature tracking sourceExperimental — accuracy not yet measured

In the paper

“Supervised methods excel in constrained settings by utilizing labeled datasets, while self-supervised approaches leverage geometric and photometric consistency to adapt effectively to unseen environments [133].”

What the sources say

• Supported

  The Supervised Learning of Visual Odometry section reports promising performance on public datasets but notes the need for labeled poses and uncertain real-world deployment. The Self-Supervised Learning of Visual Odometry section says these methods generally adapt better to new scenarios without labeled data and describes training through geometric consistency and photometric reconstruction loss (page numbers unavailable in our copy). (LLM, conf=0.90)

  cites [133] Chen, 2023, Deep learning for visual localization and mapping: A survey, IEEE Trans. Neural Netw. Learn. Syst. source

In the paper

“More advanced systems, such as LIC-Fusion [18] and its successor LIC-Fusion 2.0 [145], adopted tightly coupled optimization frameworks to deeply integrate LiDAR, IMU, and visual data, achieving a balance between high accuracy and real-time performance in large-scale environments.”

What the sources say

• Partly supported

  The Abstract and Introduction describe LIC-Fusion as a tightly coupled, efficient fusion of LiDAR, IMU, and camera measurements, and the Outdoor Tests report improved accuracy. But the paper describes an MSCKF-based filter, not a tightly coupled optimization framework, and its experiments do not establish performance in large-scale environments. LIC-Fusion 2.0 is separately cited in the citing sentence, so this verdict concerns LIC-Fusion only. (LLM, conf=0.88)

  cites [18] Zuo, 2019, Lic-fusion: Lidar-inertial-camera odometry sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [145] Shan, 2018, Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain source

In the paper

“Neural implicit representations, like NeRF [135] and Semantic-NeRF [136], encode scene geometry and semantics into compact neural fields, enabling robust localization in new environments and showcasing the potential of multimodal fusion for bridging perception and mapping.”

What the sources say

• ?Not checked — no full text of the sourcecites [135] Wang, 2023, Attention weighted local descriptors, IEEE Trans. Pattern Anal. Mach. Intell. source
• ?Could not verify

  The Method (p. 2) describes MTLDesc as a convolutional network for keypoint detection and local descriptors, not a neural field. Searches for NeRF, Semantic-NeRF, implicit representations, and multimodal fusion found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  cites [136] C. Wang, R. Xu, Y. Zhang, S. Xu, W. Meng, B. Fan, X. Zhang, MTLDesc: Looking wider to describe better, in: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 2022, pp. 2388–2396, 2. sourceExperimental — accuracy not yet measured

In the paper

“In terms of vision-IMU-GPS fusion, VINS-Fusion has become a mainstream solution for large-scale, high-precision localization and is widely used in UAVs and mobile robots.”

What the sources say

• Partly supported

  The Introduction (p. 1) describes VINS-Mono as a high-accuracy system used in drone navigation and large-scale state-estimation tasks, and the Experimental Results (p. 13) report a large-scale test. The cited paper concerns VINS-Mono, however, not VINS-Fusion; it does not establish that VINS-Fusion fuses GPS, is a mainstream solution, or is widely used in UAVs and mobile robots. (LLM, conf=0.87)

  cites [151] Qin, 2018, Vins-mono: A robust and versatile monocular visual-inertial state estimator, IEEE Trans. Robot. sourceExperimental — accuracy not yet measured

In the paper

“Additionally, the RoboMamba [153] model incorporates action dynamics into the multimodal fusion framework, optimizing vision, language, and action for task planning.”

What the sources say

• ?Could not verify

  The Conclusions describe MM3DGS as a SLAM framework combining visual, depth, and inertial measurements for rendering and trajectory tracking. A search for RoboMamba, action dynamics, and vision-language-action task planning found no passage addressing the attributed claim. (LLM, conf=0.75)

  cites [153] Sun, 2024, Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements sourceExperimental — accuracy not yet measured

In the paper

“In cross-modal self-supervised learning, tasks such as image inpainting or text mask prediction are typically designed.[183]”

What the sources say

• Partly supported

  The Introduction (p. 1) and Pre-training Objectives (p. 5) describe masked language modeling as common in vision-and-language pre-training, including prediction of masked text tokens. They identify image–text matching—not image inpainting—as the other common objective. The Conclusion (p. 9) calls for future work on visual masking objectives, so the paper does not support presenting image inpainting as a typical task. (LLM, conf=0.87)

  cites [183] Kim, 2021, Vilt: Vision-and-language transformer without convolution or region supervision sourceExperimental — accuracy not yet measured

In the paper

“For example, Mimic-Touch [161] uses a strategy learning framework that mimics human tactile-guided behaviors, collecting multimodal tactile data and combining self-supervised representation learning with reinforcement learning to optimize grasp stability.”

What the sources say

• ?Could not verify

  The Abstract (p. 1) describes 3D-VLA as a 3D vision-language-action generative world model, not a tactile-guided grasping framework. Searches for Mimic-Touch, tactile-guided behavior, self-supervised representation learning, and reinforcement learning found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  cites [161] Zhen, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Vision and language are fundamentally distinct forms of information, and cross-modal pre-training enables models to learn the deep correlations between them, making them more accurate in performing vision-language tasks [168].”

What the sources say

• Partly supported

  The “Large Vision-Language Models” section (page not provided) says these models effectively use image-text data for image-based reasoning. It does not say that vision and language are fundamentally distinct, or that cross-modal pre-training learns deep correlations that make vision-language tasks more accurate. (LLM, conf=0.82)

  cites [168] Yu, 2024 sourceExperimental — accuracy not yet measured

In the paper

“The model is trained using a contrastive loss function to bring related modality pairs, like images and descriptions, closer in a shared embedding space while distancing unrelated pairs [179,180].”

What the sources say

• Partly supported

  The Methods section (page unavailable) says CMOT projects modalities into a common manifold and minimizes distances between corresponding samples. It describes manifold alignment and optimal transport, not a contrastive loss that distances unrelated pairs; its examples concern single-cell modalities, not images and descriptions. (LLM, conf=0.88)

  cites [179] Alatkar, 2023, CMOT: cross-modality optimal transport for multimodal inference, Genome Biol. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [180] Qian, 2022, Integrating multi-label contrastive learning with dual adversarial graph neural networks for cross-modal retrieval, IEEE Trans. Pattern Anal. Mach. Intell. source

In the paper

“Similarly, VITO-Transformer [157], based on Vision Transformer (ViT), integrates tactile inputs through selfattention, enabling precise grasp point estimation and dynamic adaptability.”

What the sources say

• ?Could not verify

  The Multi-Sem Fusion Framework and AAF Module sections describe camera images, point clouds, and attention-based fusion for 3-D object detection. Searches for VITO-Transformer, tactile inputs, and grasp-point estimation found no passage addressing the attributed method or its claimed capabilities. (LLM, conf=0.76)

  cites [157] Xu, 2025 sourceExperimental — accuracy not yet measured

In the paper

“During training, the model employs specialized encoders (e.g., convolutional neural networks for visual data, transformer architectures for text and audio data) [171,172] to extract features from different modalities, and then interacts and integrates these features through cross-modal alignment mechanisms [173,174].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [171] Du, 2022
• ?Not checked — no full text of the sourcecites [172] Gan, 2022, Vision-language pre-training: Basics, recent advances, and future trends, Found. Trends® Comput. Graph. Vis. source
• Supported

  The Model Architecture section describes separate visual and text encoders used during training. Section 3.2.2 explains that cross-modal attention combines text and visual features to produce multimodal features. The paper supports this description for image and text. (LLM, conf=0.90)

  cites [173] Sun, 2023, Fusion or defusion? Flexible vision-and-language pre-training source
• Supported

  In “Global and Local Contrastive Alignment” (page unavailable), the paper says a vision encoder and a text encoder produce image and report representations that are aligned in a cross-modal space. “Report Reconstruction” (page unavailable) describes cross-attention that integrates visual and textual features. This supports the claim as attributed to this source; the source concerns images and text, while the citing sentence’s audio example is attached to other citations. (LLM, conf=0.91)

  cites [174] Zhang, 2023, Multi-task paired masking with alignment modeling for medical vision-language pre-training, IEEE Trans. Multimed. source

In the paper

“By pre-training on large-scale multimodal data, the model learns deep correlations between vision and language, significantly enhancing its ability to understand and represent information across different modalities [164,165].”

What the sources say

• ?Could not verify

  Section 3.2 describes a transformer trained on depth images, tactile contact points, joint positions, and action history for in-hand object rotation. Searches for vision-language correlations, large-scale multimodal pre-training, and cross-modal language understanding found no passage addressing the attributed claim; the copy does not provide evidence for or against it. (LLM, conf=0.76)

  cites [164] Qi, 2023, General in-hand object rotation with vision and touch sourceExperimental — accuracy not yet measured
• Partly supported

  The Broader related work section (p. 17) describes methods that use natural language for video-language alignment and a model that learns visual representations with language supervision. It does not establish that pre-training on large-scale multimodal data produces deep vision-language correlations or significantly improves cross-modal understanding, and it describes other models rather than Sparsh on this point. (LLM, conf=0.78)

  cites [165] Higuera, 2024 sourceExperimental — accuracy not yet measured

In the paper

“ALIGN [182] follows a similar approach, training cross-modal image and text data to enhance matching performance.”

What the sources say

• ?Could not verify

  The Introduction (page unavailable) describes CLIP-based methods that create a shared representation space for images and text, but it does not identify ALIGN or discuss its matching performance. Searches for “ALIGN” and “ALIGN image text matching performance” found no passage about ALIGN; passages about aligning features concern this paper’s MRI method instead. (LLM, conf=0.75)

  cites [182] Li, 2024, Joint self-supervised and supervised contrastive learning for multimodal MRI data: Towards predicting abnormal neurodevelopment, Artif. Intell. Med. sourceExperimental — accuracy not yet measured

In the paper

“Transformer was initially applied in NLP tasks, with BERT [192] and GPT [193] as key models showcasing its advantages in natural language understanding (NLU) and natural language generation (NLG), respectively.”

What the sources say

• Partly supported

  The Introduction (p. 1) describes BERT and GPT as examples of successful Transformer use in NLP, and the References (p. 10) identify BERT with language understanding. The paper calls GPT a language-modeling example, but does not explicitly describe its advantages in natural language generation or establish that NLP was the first application of Transformers. (LLM, conf=0.84)

  cites [192] M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, A. Joulin, Emerging properties in self-supervised vision transformers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9650–9660. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [193] R. Hu, A. Singh, Unit: Multimodal multitask learning with a unified transformer, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1439–1449. source

In the paper

“VQ-VAE-2 [189] improves upon traditional VAE methods and excels in multimodal generation tasks, capable of generating high-quality images or other modality content.”

What the sources say

• ?Could not verify

  The Introduction (p. 1) and Conclusion (p. 9) describe ALIGN, an image–text representation model used for retrieval, not VQ-VAE-2. Searches for VQ-VAE-2 and variational-autoencoder-based multimodal generation found no passage addressing the claimed improvement or generation capabilities. (LLM, conf=0.78)

  cites [189] Jia, 2021, Scaling up visual and vision-language representation learning with noisy text supervision sourceExperimental — accuracy not yet measured

In the paper

“Reformer [195] employed locality-sensitive hashing (LSH Attention) to reduce memory requirements for attention computation, while Longformer [196] and Linformer [197] utilized sparse attention and low-rank projections, respectively, to decrease computational complexity, making Transformer more efficient for handling long-text sequences.”

What the sources say

• ?Could not verify

  Section 3.2 (p. 5) discusses attention in a VQ-VAE-2 image model and memory constraints for its latent-code priors. Searches for Longformer, Linformer, sparse attention, and low-rank projections found no passage addressing the claimed methods for long-text sequences. (LLM, conf=0.76)

  cites [196] Razavi, 2019, Generating diverse high-fidelity images with vq-vae-2, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured
• ?Could not verify

  The Model Architecture and Pre-training Objectives sections (p. 3) discuss BLIP’s visual and text transformers, not Longformer or Linformer. Searches for those models and for long-text computational complexity found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  cites [197] Li, 2022, Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation sourceExperimental — accuracy not yet measured

In the paper

“Proposed by Google in 2018, BERT uses a Transformer encoder-only architecture and pretrains with masked language modeling (MLM) and next sentence prediction (NSP), improving text understanding tasks.”

What the sources say

• Partly supported

  The Introduction (p. 1) identifies BERT as an example of self-supervised pretraining and says such pretraining contributed to Transformers’ success in NLP. It does not specify that BERT is encoder-only, name MLM or NSP, or describe results on text-understanding tasks. The implementation discussion (p. 4) describes a Vision Transformer, not BERT’s architecture. (LLM, conf=0.86)

  cites [192] M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, A. Joulin, Emerging properties in self-supervised vision transformers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9650–9660. sourceExperimental — accuracy not yet measured

In the paper

“Since its introduction by Vaswani et al. [191] in 2017, Transformer has become a core architecture in deep learning, achieving major advancements in natural language processing (NLP), computer vision (CV), and multimodal learning.”

What the sources say

• ?Could not verify

  The Introduction (p. 1) and Conclusion (p. 13) discuss CNNs and cross-modal self-supervision for visual recognition, not Transformer. Searches for Transformer, Vaswani, self-attention, and NLP found no passage in our copy addressing the claimed architecture or its advances across those fields. (LLM, conf=0.76)

  cites [191] Sayed, 2019, Cross and learn: Cross-modal self-supervision sourceExperimental — accuracy not yet measured

In the paper

“UniT [186], on the other hand, is a self-supervised learning framework for cross-modal tasks.It jointly pre-trains different modalities (e.g., vision and text) through self-supervised tasks, enabling the model to learn universal features.”

What the sources say

• ?Could not verify

  The Abstract (p. 1) describes a cross-modal contrastive approach for learning music-audio features from playlists, genre metadata, and audio. Searches for UniT and for joint self-supervised pre-training of vision and text found no passage addressing the claimed framework or its universal features. (LLM, conf=0.77)

  cites [186] Ferraro, 2021, Enriched music representations with multiple cross-modal contrastive learning, IEEE Signal Process. Lett. sourceExperimental — accuracy not yet measured

In the paper

“DeepSeek-MoE [21], using 64 experts and dynamic activation, drastically reduces computational redundancy while maintaining high performance, and improves expert scheduling for better generalization in multilingual and multi-task settings.”

What the sources say

• Partly supported

  Section 4.2 reports 64 experts (one shared and 63 routed), and Section 3.2 describes routing that activates selected experts. The Conclusion reports performance comparable to two dense models with about 40% of their computation, while Section 3.2 addresses redundancy among expert parameters. The paper does not establish improved expert scheduling or better generalization specifically in multilingual and multi-task settings; it also does not describe the measured computation saving as a reduction in computational redundancy. (LLM, conf=0.88)

  cites [21] Dai, 2024 sourceExperimental — accuracy not yet measured

In the paper

“MoE has been integrated into ultra-large-scale models like GPT-4 [22] and DeepSeek-MoE [21], advancing multi-task and multimodal applications.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [22] Achiam, 2023
• Partly supported

  The Introduction and Section 5 identify DeepSeekMoE as a MoE language model scaled to 16B parameters. Section 6 mentions multimodal models built with other MoE architectures, but the paper does not present DeepSeekMoE as advancing multimodal applications or establish the claimed multi-task applications. The GPT-4 portion has a separate citation and is not assessed here. (LLM, conf=0.85)

  cites [21] Dai, 2024 sourceExperimental — accuracy not yet measured

In the paper

“GLaM [202] refined MoE by introducing adaptive expert selection, dynamically routing inputs to the most relevant experts, balancing efficiency and performance.”

What the sources say

• ?Could not verify

  The Related Work section (p. 3) discusses language models used for embodied planning, while the Summary of Experiments (p. 9) discusses PaLM-E’s transfer and data efficiency. Neither addresses GLaM or mixture-of-experts routing; searches for adaptive expert selection and sparse expert routing found no relevant passage in our copy. (LLM, conf=0.75)

  cites [202] Driess, 2023 sourceExperimental — accuracy not yet measured

In the paper

“For instance, the FLAN dataset [214] covers thousands of task instructions across multiple domains, allowing models to generalize effectively.”

What the sources say

• ?Could not verify

  The Experiments section (p. 8) describes DETR experiments on the COCO object-detection and segmentation datasets, not the FLAN dataset. Searches for FLAN, task instructions, and instruction tuning found no passage addressing the claimed dataset size, domain coverage, or model generalization. (LLM, conf=0.78)

  cites [214] Carion, 2020, End-to-end object detection with transformers sourceExperimental — accuracy not yet measured

In the paper

“LLaVA [211]'s ability to adjust to different tasks using natural language instructions enables it to not only respond rapidly to traditional visuallanguage tasks like VQA and image captioning, but also to adjust its execution strategy dynamically for new tasks.”

What the sources say

• ?Could not verify

  The Introduction (p. 2) describes Linformer’s pretraining and evaluation on language tasks, and the Model section (p. 4) describes its linear self-attention mechanism. Searches for LLaVA, natural-language instructions, VQA, image captioning, and dynamic task strategies found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  cites [211] Wang, 2020 sourceExperimental — accuracy not yet measured

In the paper

“InternVL [212] further optimizes efficiency through expert scheduling and distributed computing, ensuring dynamic resource allocation for large-scale multimodal data.”

What the sources say

• ?Could not verify

  The scaling study (Section 4.4, p. 8) and computational-cost analysis (Appendix D.5, p. 18) discuss Vision Transformer efficiency, not InternVL. Searches for InternVL, expert scheduling, distributed computing, and dynamic resource allocation found no passage addressing the attributed claim. (LLM, conf=0.78)

  cites [212] Dosovitskiy, 2020 sourceExperimental — accuracy not yet measured

In the paper

“LLaVA [211] also uses similar strategies to generate more natural, user-friendly responses.”

What the sources say

• ?Could not verify

  The Introduction (p. 1) discusses efficient self-attention, and Downstream Results (p. 7) reports Linformer evaluations. Searches for LLaVA, RLHF, instruction tuning, and user-friendly responses found no passage in our copy addressing the claim. (LLM, conf=0.75)

  cites [211] Wang, 2020 sourceExperimental — accuracy not yet measured

In the paper

“Through instruction tuning, LLaVA [211] can generate precise image descriptions or answer visual-related questions based on different task demands.”

What the sources say

• ?Could not verify

  The Introduction (p. 2) and Downstream Results (p. 7) discuss Linformer’s performance on language tasks, not LLaVA. Searches for LLaVA, instruction tuning, image descriptions, and visual question answering found no passage in our copy addressing the claim. (LLM, conf=0.77)

  cites [211] Wang, 2020 sourceExperimental — accuracy not yet measured

In the paper

“MoE-LLaVA [208] uses a Mixture of Experts (MoE) architecture, activating only the relevant expert networks to reduce computational burden and improve reasoning efficiency, especially on large datasets.”

What the sources say

• ?Could not verify

  The Introduction (pp. 3–4) discusses scaling transformer language models and in-context learning, not MoE-LLaVA. A mixture-of-experts paper appears in the References (p. 74), but our copy does not discuss MoE-LLaVA’s architecture, expert activation, computational burden, or reasoning efficiency (searched: MoE-LLaVA and mixture-of-experts efficiency). (LLM, conf=0.78)

  cites [208] Brown, 2020, Language models are few-shot learners, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“Context-aware modeling faces several challenges, such as handling dynamically changing contexts [223], balancing the significance of different modality data [224], and preventing overfitting.”

What the sources say

• ?Could not verify

  The Discussion (p. 9) concerns instruction tuning and generalization to unseen tasks, not dynamically changing contexts or the relative importance of different modalities. Appendix B.1 (p. 24) mentions overfitting to instruction templates, but does not address it as a challenge of context-aware modeling. Searches for the context and modality claims found no supporting passage in our copy. (LLM, conf=0.75)

  cites [223] Wei, 2021 sourceExperimental — accuracy not yet measured
• ?Could not verify

  Citation [223] is attached to the challenge of handling dynamically changing contexts. The Related Work section discusses multitask NLP prompting, and the retrieved QA prompt templates use “context” as an input field; neither addresses contexts changing dynamically. Searches for dynamically changing and adapting to changing contexts found no supporting passage. (LLM, conf=0.72)

  cites [224] Sanh, 2021 sourceExperimental — accuracy not yet measured

In the paper

“To address these, edge computing has emerged to process data directly at the network edge [229], handling both downstream cloud data and upstream IoT data.”

What the sources say

• Partly supported

  Section 3.1 (System Architecture; page unavailable) describes training user-specific models locally on edge devices, which supports local processing at the edge. It does not describe edge computing’s emergence or say that it handles downstream cloud data and upstream IoT data; searches for those terms found no supporting passage. (LLM, conf=0.78)

  cites [229] Malikireddy, 2024 sourceExperimental — accuracy not yet measured

In the paper

“In VQA, models like GPT-4V provide accurate responses in diverse domains, and LLaVA [211] enhances visual understanding for detailed answers.”

What the sources say

• ?Could not verify

  The citation is attached to the claim about LLaVA. The Introduction (p. 1) discusses NLP transformers and question answering, while the downstream results (p. 7) concern text-based tasks. Searches for LLaVA, visual understanding, and visual question answering found no passage addressing LLaVA’s ability to provide detailed answers. (LLM, conf=0.76)

  cites [211] Wang, 2020 sourceExperimental — accuracy not yet measured

In the paper

“For example, BLIP-2 employs the Query Transformer to optimize vision-language information extraction and adapts to instructions for image captioning and VQA.”

What the sources say

• Partly supported

  The Vision Encoder section identifies the Q-former as BLIP-2’s method for connecting image and language representations, and Table 3 shows BLIP-2 describing an image. But Table 3 says BLIP-2 focuses on describing the image rather than following the user’s question instruction; the copy does not support the claimed adaptation to instructions for VQA. (LLM, conf=0.88)

  cites [218] Liu, 2024, Visual instruction tuning, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“Similarly, T0 [215] trains language models using a large-scale instruction dataset, endowing them with zero-shot and few-shot learning capabilities.”

What the sources say

• ?Could not verify

  The Abstract (p. 1) and Downstream Results (p. 14) describe Switch Transformers, not T0. Searches for T0, instruction datasets, and zero-shot or few-shot learning found no passage in our copy addressing the attributed claim. (LLM, conf=0.77)

  cites [215] Fedus, 2022, Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, J. Mach. Learn. Res. sourceExperimental — accuracy not yet measured

In the paper

“PaLM-E [204] integrates language, vision, and robotic control signals, optimizing reasoning performance in physical-world tasks.”

What the sources say

• ?Could not verify

  The Introduction and Related Work sections discuss Qwen-VL and other vision-language models, but do not describe PaLM-E. Searches for PaLM-E, robotic control signals, and embodied robotic reasoning found no passage addressing its claimed integration or physical-world reasoning performance; page numbers were not available in our copy. (LLM, conf=0.76)

  cites [204] Bai, 2023 sourceExperimental — accuracy not yet measured

In the paper

“Context-aware modeling is widely used across various domains, including natural language processing [219], computer vision [220], and recommendation systems [221], where it can substantially boost both system performance and user experience.”

What the sources say

• ?Not checked — no full text of the sourcecites [219] Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu, et al., Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 24185–24198. source
• ?Could not verify

  The Related Work section says large language models have advanced natural language processing, but it does not say that context-aware modeling is widely used there or that it boosts performance and user experience. Searches for “context-aware modeling” and “context-aware” found no passage addressing that point. The other domains in the sentence have separate citations. (LLM, conf=0.72)

  cites [220] Lu, 2024 sourceExperimental — accuracy not yet measured
• Partly supported

  Section 4.3 (page not available in our copy) describes conditioning a vision-language model on a question to focus on informative image regions, and Section 4.1 reports strong visual-question-answering performance. This supports an example of context-dependent modeling in computer vision, the domain attached to this citation, but does not establish that it is widely used or that it improves user experience. (LLM, conf=0.82)

  cites [221] Li, 2023, Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models sourceExperimental — accuracy not yet measured

In the paper

“Kosmos-2 [210] strengthens the joint modeling of images and language through self-supervised learning, allowing it to handle more complex tasks, such as deep reasoning and generation of image and text.”

What the sources say

• ?Could not verify

  The Introduction (p. 1) describes Longformer as a model for processing long text sequences, and Related Work (p. 2) discusses long-document transformers. Searches for Kosmos-2, joint image–language modeling, and image-and-text generation found no passage addressing the attributed claim. (LLM, conf=0.77)

  cites [210] Beltagy, 2020 sourceExperimental — accuracy not yet measured

In the paper

“LLaVA [211] (Language and Vision Assistant) is an innovative visual-language model that employs instruction tuning, giving the model a high degree of flexibility in handling various tasks.”

What the sources say

• ?Could not verify

  The Introduction (p. 2) describes pretraining and fine-tuning Linformer on language tasks, not LLaVA or visual-language instruction tuning. Searches for LLaVA and for visual-language instruction tuning found no passage addressing the claimed task flexibility. (LLM, conf=0.78)

  cites [211] Wang, 2020 sourceExperimental — accuracy not yet measured

In the paper

“In robotic navigation, PaLM-E [204] predicts optimal paths, avoids obstacles, and provides action recommendations.”

What the sources say

• ?Could not verify

  The Introduction and Related Work discuss Qwen-VL and other vision-language models, but do not describe PaLM-E performing robotic navigation. Searches for PaLM-E, path planning, obstacle avoidance, and robot actions found no passage supporting the attributed capabilities. (LLM, conf=0.76)

  cites [204] Bai, 2023 sourceExperimental — accuracy not yet measured

In the paper

“Llama [209] uses large-scale pretraining to boost text understanding and generation, improving performance in tasks like visual reasoning and image-to-text generation when combined with visual data.”

What the sources say

• ?Could not verify

  The Introduction (p. 2) reports Reformer experiments on text and image generation separately, while Related Work (p. 7) discusses large language models generally. A search for Llama, visual reasoning, and image-to-text generation found no passage addressing the attributed claim. (LLM, conf=0.76)

  cites [209] Kitaev, 2020 sourceExperimental — accuracy not yet measured

In the paper

“Gemini [11] further extends this approach by jointly training largescale data from images, text, and audio, enhancing the model's cross-modal reasoning ability, particularly excelling in visual reasoning tasks.”

What the sources say

• Supported

  Section 5 says Gemini is trained jointly across text, image, audio, and video and reports strong performance across those domains. The Introduction describes cross-modal reasoning and reports state-of-the-art results for Gemini Ultra on image-understanding benchmarks and the image-based MMMU reasoning benchmark. (LLM, conf=0.90)

  cites [11] Team, 2023 source

In the paper

“Additionally, PaLM-E [204] combines improvements in multimodal learning and reasoning capabilities, enabling it to process multimodal data such as images, text, and audio, performing well in cross-modal reasoning and generation tasks.”

What the sources say

• ?Could not verify

  The Introduction and Related Work discuss Qwen-VL and other vision-language models, but do not describe PaLM-E. Searches for “PaLM-E” and for image, text, audio, and cross-modal reasoning capabilities found no passage supporting the attributed claim. (LLM, conf=0.78)

  cites [204] Bai, 2023 sourceExperimental — accuracy not yet measured

In the paper

“Ovis1.6 [213] balances computational efficiency and reasoning accuracy, making it ideal for real-time visual-language tasks.”

What the sources say

• ?Could not verify

  The Introduction describes Swin Transformer as a computer-vision backbone, and the Ablation Study reports speed improvements for Swin models. Searches for Ovis1.6 and real-time visual-language reasoning found no passage addressing that model or its claimed balance of efficiency and accuracy. (LLM, conf=0.78)

  cites [213] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10012–10022. sourceExperimental — accuracy not yet measured

In the paper

“nuScenes [23] offers data from six cameras, five radars, and one LiDAR, covering 360° views and supporting 3D detection and tracking tasks across 23 categories.”

What the sources say

• Supported

  Section 2, Table 2 lists six cameras, five radars and one lidar, with 360° camera coverage. Section 2 also reports annotations for 23 object classes, and the Conclusion identifies detection and tracking as nuScenes tasks. Page numbers are not provided in our copy. (LLM, conf=0.94)

  cites [23] H. Caesar, V. Bankiti, A.H. Lang, S. Vora, V.E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, O. Beijbom, nuscenes: A multimodal dataset for autonomous driving, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11621–11631. source

In the paper

“The former covers 564 scenes and 86 tasks, combining multi-view camera streams and language annotations to enhance robots' adaptability to complex scenes.”

What the sources say

• Supported

  Table I lists DROID with 564 scenes, 86 verbs and language instructions; the Discussion explains that the paper uses unique instruction verbs to count tasks. The robot-platform description identifies multiple camera views, and the Experiments report that training with DROID improves policy robustness to scene changes. (LLM, conf=0.91)

  cites [25] Khazatsky, 2024 source

In the paper

“NYU Depth V2 [239], as one of the earliest RGB-D datasets, offers 1,449 images from 464 diverse indoor scenes with detailed semantic segmentation and object relationship annotations, making it instrumental for research on physical support reasoning and cluttered environments.”

What the sources say

• ?Could not verify

  The Mobile Semantic Segmentation section (p. 8) describes experiments on PASCAL VOC 2012, and the unseen-dataset section (p. 23) discusses MS-COCO. Searches for NYU Depth V2, the 1,449-image and 464-scene figures, and object-relationship annotations found no passage in our copy addressing the claim. (LLM, conf=0.78)

  cites [239] Mehta, 2021 sourceExperimental — accuracy not yet measured

In the paper

“SimpleBEV [260], adopting a BEV-based multi-scale fusion strategy, strengthens depth perception and maintains stable overall detection performance.”

What the sources say

• ?Could not verify

  The Abstract and §5.2 describe REVERIE, a navigation and object-identification task, rather than SimpleBEV or its detection performance. Searches for SimpleBEV, BEV-based multi-scale fusion, and bird’s-eye-view detection found no passage in our copy addressing the attributed claim. (LLM, conf=0.78)

  cites [260] Y. Qi, Q. Wu, P. Anderson, X. Wang, W.Y. Wang, C. Shen, A.v.d. Hengel, Reverie: Remote embodied visual referring expression in real indoor environments, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 9982–9991. sourceExperimental — accuracy not yet measured

In the paper

“Unlike the aforementioned datasets, R3ED [251] focuses on real-world data, providing 5,800 point clouds and 22,400 object annotations collected using IntelRealsense D455 sensors.It supports active visual learning tasks, such as next-best-view planning and 3D object detection.”

What the sources say

• ?Could not verify

  Section 4 and Appendix A.3 describe R2Sim, a simulated grasp dataset, not R3ED. Section 5.3 mentions an Intel RealSense D455 camera for grasping experiments, but our copy does not address R3ED, its claimed counts, or its use for active visual learning (searched: R3ED, 5,800, 22,400, and next-best-view planning). (LLM, conf=0.77)

  cites [251] Cai, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Building upon Matterport3D [26], Room-to-Room (R2R) [251] was the first dataset to integrate natural language modalities, introducing the Vision-and-Language Navigation (VLN) task.”

What the sources say

• ?Could not verify

  The cited paper’s Section 4 describes R2Sim, a simulated dataset for grasp detection, not Room-to-Room. Searches for Room-to-Room, Matterport3D, natural-language instructions, and Vision-and-Language Navigation found no passage addressing the claim. The copy is marked complete, but its silence does not establish a contrary finding. (LLM, conf=0.78)

  cites [251] Cai, 2024 sourceExperimental — accuracy not yet measured

In the paper

“On the other hand, lightweight natural language processing (NLP) models such as TinyBERT [232] focus on reducing the size of transformer-based models via knowledge distillation, making them suitable for mobile deployment.”

What the sources say

• ?Could not verify

  The Introduction and Related Works sections describe MulT, a transformer for unaligned multimodal sequences, rather than TinyBERT. Searches for TinyBERT, knowledge distillation, model compression, and mobile deployment found no passage addressing the attributed claim; page numbers were not available in our copy. (LLM, conf=0.76)

  cites [232] Y.-H.H. Tsai, S. Bai, P.P. Liang, J.Z. Kolter, L.-P. Morency, R. Salakhutdinov, Multimodal transformer for unaligned multimodal language sequences, in: Proceedings of the Conference. Association for Computational Linguistics. Meeting, vol. 2019, 2019, p. 6558. sourceExperimental — accuracy not yet measured

In the paper

“Zhang et al. proposed a non-intrusive appliance state detection system [235] that uses a single sensor to monitor multiple appliances via a cloud-edgedevice collaborative framework.”

What the sources say

• ?Could not verify

  The Abstract and Introduction describe a method for learning representations of multi-word expressions, not an appliance-monitoring system. Searches for non-intrusive appliance detection, a single sensor, and a cloud-edge-device framework found no passage addressing the claim. (LLM, conf=0.78)

  cites [235] R. Vacareanu, M.A. Valenzuela-Escarcega, R. Sharp, M. Surdeanu, An unsupervised method for learning representations of multi-word expressions for semantic classification, in: Proceedings of the 28th International Conference on Computational Linguistics, 2020, pp. 3346–3356. sourceExperimental — accuracy not yet measured

In the paper

“In addition, the Hypersim dataset [240], utilizing high-fidelity synthetic techniques, generates 77,400 indoor scene images with detailed pixel-level annotations, light decomposition, and surface normals, serving as an important tool for simulation-to-reality (Sim-to-Real) transfer research.”

What the sources say

• ?Could not verify

  The TinyBERT Settings and GLUE Details sections discuss a distilled language model and natural-language benchmarks, not the Hypersim dataset. Searches for Hypersim, 77,400 images, synthetic indoor scenes, surface normals, and simulation-to-reality transfer found no passage addressing the claim. (LLM, conf=0.78)

  cites [240] Jiao, 2019 sourceExperimental — accuracy not yet measured

In the paper

“ScanNet [238], through its 2.5 million RGB-D frames across 1,513 scenes, utilizes crowdsourced annotations to provide a large-scale, high-quality dataset for semantic segmentation and 3D modeling.”

What the sources say

• ?Could not verify

  The transfer-learning dataset table (Section 5.2, p. 8) lists image-classification datasets, not ScanNet. Searches for ScanNet, 2.5 million RGB-D frames, 1,513 scenes, and semantic segmentation found no passage in our copy addressing the claimed dataset or its annotations. (LLM, conf=0.78)

  cites [238] Tan, 2019, Efficientnet: Rethinking model scaling for convolutional neural networks sourceExperimental — accuracy not yet measured

In the paper

“Early sequence-based models like Seq2Seq [265] and SF [266] struggled with global path optimization, leading to poor performance in unseen environments.”

What the sources say

• ?Not checked — no full text of the sourcecites [265] H. Liu, Y. Teng, T. Lu, H. Wang, L. Wang, Sparsebev: High-performance sparse 3d object detection from multi-camera videos, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 18580–18590. source
• ?Could not verify

  Section 4.1 (Dataset and Metrics; page unavailable) identifies nuScenes and Argoverse 2 as the paper’s evaluation datasets, not R2R. Searches for Seq2Seq, SF, and global path optimization found no passage addressing navigation performance in unseen environments. (LLM, conf=0.78)

  cites [266] Liu, 2024, Ray denoising: Depth-aware hard negative sampling for multi-view 3d object detection sourceExperimental — accuracy not yet measured

In the paper

“R2R [251] provides 21,567 crowd-sourced navigation instructions paired with panoramic RGB-D data, enabling agents to complete path navigation following language commands.”

What the sources say

• ?Could not verify

  Section 4, “R2Sim Dataset,” describes simulated RGB-D images and grasp annotations, not the R2R navigation dataset. Searches for “21,567 crowd-sourced navigation instructions” and “R2R Room-to-Room navigation” found no passage addressing the claimed instructions or language-guided navigation. Our copy therefore cannot verify what the citing sentence attributes to this source. (LLM, conf=0.77)

  cites [251] Cai, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Lu et al. introduced CLONE [234], a framework for cooperative learning across edge devices, using real-world data from an electric vehicle company to focus on battery and component failures.”

What the sources say

• ?Could not verify

  The cited paper’s discussion of multimodal models and reinforcement learning from AI feedback, and its References section, do not address CLONE. Searches of our full-text copy for cooperative learning across edge devices and for electric-vehicle battery or component failures found no relevant passage. (LLM, conf=0.77)

  cites [234] Ahn, 2024 sourceExperimental — accuracy not yet measured

In the paper

“To address this, Zhang et al. proposed a cloud-edge collaborative system [233], which uses data like vehicle position, speed, and acceleration to model driving behavior and detect abnormal actions like speeding or drunk driving.”

What the sources say

• ?Could not verify

  Our copy’s Conclusion describes CrisisSpot as a method for classifying disaster-related social media content, not a vehicle-behavior system. Searches for cloud-edge collaboration, vehicle position, speed, acceleration, speeding, and drunk driving found no passage addressing the attributed claim. (LLM, conf=0.76)

  cites [233] Dar, 2025, A social context-aware graph-based multimodal attentive learning framework for disaster content classification during emergencies, Expert Syst. Appl. sourceExperimental — accuracy not yet measured

In the paper

“Additionally, SemanticKITTI [24] focuses on semantic segmentation of LiDAR point clouds and provides an important benchmark for sequence-level semantic scene completion.”

What the sources say

• Supported

  Section 5, “Evaluation of Semantic Scene Completion” (page unavailable), describes semantic point-cloud segmentation using sequences of past scans and presents SemanticKITTI as a real-world outdoor benchmark for 3D semantic scene completion. It explains that the dataset’s sequential scans make the completion benchmark possible. (LLM, conf=0.91)

  cites [24] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, J. Gall, Semantickitti: A dataset for semantic scene understanding of lidar sequences, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 9297–9307. source

In the paper

“However, due to the limitations of Cop-peliaSim's physics engine, RLBench [288] remains restricted to purely simulated tasks, posing challenges for sim-to-real transfer.”

What the sources say

• ?Could not verify

  Section 4.1 describes experiments on the R2R vision-language navigation dataset, and Section 4.2 describes R4R; neither addresses RLBench. Searches for RLBench, CoppeliaSim physics limitations, and sim-to-real transfer found no passage in our copy that speaks to the claim. (LLM, conf=0.76)

  cites [288] H. Wang, W. Wang, W. Liang, C. Xiong, J. Shen, Structured scene memory for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8455–8464. sourceExperimental — accuracy not yet measured

In the paper

“EnvEdit [279], through environment augmentation, reduces overfitting and enhances robustness in unseen scenarios.”

What the sources say

• Partly supported

  Section 4.2.1 (p. 7) reports that speaker-driven data augmentation improved success on unseen validation environments, supporting the generalization part of the claim. But Section 4.1 (p. 6) describes augmenting routes sampled from training environments, not environment augmentation; our copy does not identify the method as EnvEdit or report a reduction in overfitting. (LLM, conf=0.86)

  cites [279] Fried, 2018, Speaker-follower models for vision-and-language navigation, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“While GemBench [289] presents unique advantages in multimodal fusion and generalization, it also remains constrained to simulation and lacks real-world validation.”

What the sources say

• ?Could not verify

  The Experimental Results (Section 5) and Conclusion (Section 6) discuss Airbert’s performance on vision-and-language navigation tasks, not GemBench. A search for GemBench and for simulation or real-world validation found no passage in our copy assessing GemBench’s advantages or limitations. (LLM, conf=0.78)

  cites [289] P.-L. Guhur, M. Tapaswi, S. Chen, I. Laptev, C. Schmid, Airbert: In-domain pretraining for vision-and-language navigation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1634–1643. sourceExperimental — accuracy not yet measured

In the paper

“RLBench [288] is one of the most widely used robotic manipulation benchmarks, offering 100+ tasks ranging from basic object grasping to complex tool use and assembly tasks.”

What the sources say

• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [288] H. Wang, W. Wang, W. Liang, C. Xiong, J. Shen, Structured scene memory for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8455–8464. sourceExperimental — accuracy not yet measured

In the paper

“KitchenShift [292] is particularly useful for benchmarking imitation learning (IL) and reinforcement learn-ing (RL) strategies, but like RLBench [288], it does not support real-world reproducibility, though its domain shift testing provides insight into real-world uncertainty modeling.”

What the sources say

• ?Not checked — no full text of the sourcecites [292] J. Li, H. Tan, M. Bansal, Envedit: Environment editing for vision-and-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 15407–15417. source
• ?Could not verify

  Sections 4.1 and 4.2 evaluate vision-language navigation on R2R and R4R, not KitchenShift. Searches for KitchenShift, domain shift, and real-world reproducibility found no passage addressing the claimed benchmark or its properties; the paper’s mention of a kitchen is a navigation example. (LLM, conf=0.78)

  cites [288] H. Wang, W. Wang, W. Liang, C. Xiong, J. Shen, Structured scene memory for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8455–8464. sourceExperimental — accuracy not yet measured

In the paper

“Chat-Scene [286] and Scene-LLM [285] optimize structured embeddings and object-based feature representations, excelling in object-related queries and scene reasoning tasks.”

What the sources say

• ?Not checked — no full text of the sourcecites [286] D. An, Y. Qi, Y. Huang, Q. Wu, L. Wang, T. Tan, Neighbor-view enhanced model for vision and language navigation, in: Proceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 5101–5109. source
• ?Could not verify

  The Methodology (p. 3) describes a vision-language navigation agent, and the Experimental Setup (p. 11) describes viewpoint and instruction embeddings. Searches for Chat-Scene, Scene-LLM, object-based features, and object-related queries found no passage in our copy addressing the attributed claim. (LLM, conf=0.77)

  cites [285] Wang, 2020, Active visual information gathering for vision-language navigation sourceExperimental — accuracy not yet measured

In the paper

“However, as RLBench is limited to simulation, it lacks real-world reproducibility, restricting its applicability in sim-to-real transfer.”

What the sources say

• ?Could not verify

  The Results sections (§§4.1–4.2; page numbers unavailable) evaluate vision-language navigation on R2R and R4R, not RLBench. Searches for RLBench, simulation-only limitations, real-world reproducibility, and sim-to-real transfer found no passage addressing the claim. (LLM, conf=0.76)

  cites [288] H. Wang, W. Wang, W. Liang, C. Xiong, J. Shen, Structured scene memory for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8455–8464. sourceExperimental — accuracy not yet measured

In the paper

“3D-VLP [281] employs a context-aware alignment strategy, performing well”

What the sources say

• ?Could not verify

  The Introduction (p. 1) and Visual and Textual Co-Grounding section (p. 3) describe a self-monitoring vision-and-language navigation agent, not 3D-VLP. A search for 3D-VLP, context-aware alignment, and dense scene description found no passage in our copy addressing the attributed claim. (LLM, conf=0.78)

  cites [281] Ma, 2019 sourceExperimental — accuracy not yet measured

In the paper

“BEVBert [116], integrating hybrid topometric pretraining, achieves state-of-the-art SR and SPL, demonstrating superior generalization across seen and unseen environments.”

What the sources say

• ?Could not verify

  The Comparison with State-of-the-Arts section (pp. 5–6) reports SR and SPL for InstruGen and other named methods, but does not discuss BEVBert. Searches for BEVBert, hybrid topometric pretraining, and its seen/unseen results found no passage addressing the attributed claim. (LLM, conf=0.76)

  cites [116] Yan, 2024 sourceExperimental — accuracy not yet measured

In the paper

“PRET (DINOv2) [119] and other vision-language pretraining (VLP) methods improve cross-modal alignment, with PRET (DINOv2) achieving higher SR and SPL through self-supervised learning.”

What the sources say

• Partly supported

  The Visual Representation in Vision-Language Pre-training discussion says grid-based pre-training enables multi-grained vision-language alignments, and the Conclusion (p. 8) says BEVBert enhances spatial-aware cross-modal reasoning. The paper reports higher SR and SPL for BEVBert, not PRET (DINOv2); the searched passages do not support the claim that PRET achieves those results through self-supervised learning. (LLM, conf=0.88)

  cites [119] An, 2022 sourceExperimental — accuracy not yet measured

In the paper

“RLBench [288] is built on PyRep (CoppeliaSim), a simulator that provides a diverse set of robotic manipulation tasks with high-dimensional visual inputs, making it well-suited for endto-end policy learning.”

What the sources say

• ?Could not verify

  The paper’s §§4.1–4.2 discuss vision-language navigation on R2R and R4R, while §3.4 describes training a navigation agent. Searches for RLBench, PyRep, CoppeliaSim, manipulation tasks, and end-to-end policy learning found no passage addressing the attributed description. (LLM, conf=0.78)

  cites [288] H. Wang, W. Wang, W. Liang, C. Xiong, J. Shen, Structured scene memory for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8455–8464. sourceExperimental — accuracy not yet measured

In the paper

“Among them, LLaVA-3D [280] and 3D-LLM [282] demonstrate strong generalization capabilities, excelling in question-answering accuracy and spatial reasoning, primarily due to their optimized multi-view feature extraction and 3D spatial embedding mechanisms.”

What the sources say

• ?Not checked — no full text of the sourcecites [280] X. Wang, Q. Huang, A. Celikyilmaz, J. Gao, D. Shen, Y.-F. Wang, W.Y. Wang, L. Zhang, Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 6629–6638. source
• ?Could not verify

  The supplied passages contain bibliographic metadata for a paper on navigation in unseen environments, not findings about ScanQA, LLaVA-3D, or 3D-LLM. Searches of our indexed copy for those names and for question-answering accuracy, spatial reasoning, multi-view features, and 3D embeddings returned no passage addressing the claim. (LLM, conf=0.76)

  cites [282] Tan, 2019 sourceExperimental — accuracy not yet measured

In the paper

“In contrast, GemBench [289], built upon RLBench, focuses on task generalization, incorporating zero-shot task setups and vision-language extensions to assess robotic performance in previously unseen tasks and objects.”

What the sources say

• ?Could not verify

  The Introduction and Section 3 describe Airbert and its BnB dataset for vision-and-language navigation. Section 5 evaluates navigation on R2R and REVERIE, not GemBench. Searches for GemBench, RLBench, zero-shot task setups, and unseen robotic tasks or objects found no passage addressing the attributed description. (LLM, conf=0.78)

  cites [289] P.-L. Guhur, M. Tapaswi, S. Chen, I. Laptev, C. Schmid, Airbert: In-domain pretraining for vision-and-language navigation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1634–1643. sourceExperimental — accuracy not yet measured

In the paper

“Current mainstream VLMs [209,206,204,208] are usually pre-trained based on large-scale Transformer architectures, which usually have billions or even tens of billions of parameters, and pre-training on large-scale datasets requires a large amount of computational resources and time costs, which limits the training efficiency and scale-up ability of the models, and makes the model deployment face”

What the sources say

• Partly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [209] Kitaev, 2020 sourceExperimental — accuracy not yet measured
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [206] Lin, 2024
• Partly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [204] Bai, 2023 sourceExperimental — accuracy not yet measured
• Partly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [208] Brown, 2020, Language models are few-shot learners, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“The challenges of low-quality data are mainly reflected in various aspects such as noise pollution, missing modalities, modal quality imbalance, and dynamic changes in modal quality, etc., which are widely found in many application scenarios such as autonomous driving [81,88,257], medical image analysis, speech recognition [296], etc., and pose great challenges to the validity and reliability of”

What the sources say

• Supported

  This citation is attached to the autonomous-driving example, not the later medical-image or speech-recognition examples. The Experimental Setup (p. 6) identifies the dataset as an autonomous-driving benchmark, and Main Results (p. 9) examines camera loss and sensor jitter as sensor errors during inference. (LLM, conf=0.90)

  cites [81] Li, 2022, Unifying voxel-based representation with transformer for 3d object detection, Adv. Neural Inf. Process. Syst. source
• Partly supported

  The Introduction (pp. 1–2) identifies autonomous driving as an application of camera–LiDAR detection and describes limitations of each sensor. It also reports that camera data compensate for lower-resolution LiDAR data (p. 4). This supports an autonomous-driving example of uneven sensor quality, but the paper does not establish that the full list of low-quality-data challenges occurs in that setting. (LLM, conf=0.82)

  cites [88] Yoo, 2020, 3D-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [257] A.D. Vuong, M.N. Vu, B. Huang, N. Nguyen, H. Le, T. Vo, A. Nguyen, Language-driven Grasp Detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17902–17912. source
• ?Could not verify

  The Limitation and Discussion section (p. 11) describes blurriness and geometric inconsistencies in generated panoramas, while the Comparisons section (p. 8) mentions Cityscapes only as a dataset not used for training. Searches for the listed low-quality-data challenges and their occurrence in autonomous driving, medical image analysis, or speech recognition found no passage addressing that claim. (LLM, conf=0.76)

  cites [296] Wang, 2025, PanoGen++: Domain-adapted text-guided panoramic environment generation for vision-and-language navigation, Neural Netw. sourceExperimental — accuracy not yet measured

In the paper

“However, existing large-scale visual language datasets (e.g., Visual Genome [311], Conceptual Captions [312], and LAION-400m [313]) generally suffer from limited size, high annotation noise, insufficient semantic diversity, and expensive annotation.”

What the sources say

• ?Could not verify

  The References (p. 12) mention Visual Genome, but do not assess its size, annotation noise, semantic diversity, or annotation cost. The Baselines (p. 6) discuss robotic-manipulation models rather than these datasets; searches for the named datasets and their alleged limitations found no passage addressing the claim. (LLM, conf=0.78)

  cites [311] Pumacay, 2024 sourceExperimental — accuracy not yet measured
• ?Could not verify

  The Introduction (p. 2) describes VIMA’s robot-manipulation model, and the Evaluation Results (pp. 6–7) discuss its performance on VIMA-BENCH. Searches for the named visual-language datasets and for annotation noise, semantic diversity, and annotation cost found no passage in our copy assessing those datasets or stating the claimed limitations. (LLM, conf=0.75)

  cites [312] Jiang, 2022 sourceExperimental — accuracy not yet measured
• ?Could not verify

  The Introduction (pp. 2–3) discusses generating robotic simulation tasks and reducing labeling effort for simulated demonstrations, not the properties of visual-language datasets. Searches for Visual Genome, Conceptual Captions, LAION-400m, annotation noise, and semantic diversity found no passage assessing the named datasets or the claimed limitations. (LLM, conf=0.75)

  cites [313] Hua, 2024 sourceExperimental — accuracy not yet measured

In the paper

“e.g., in multispectral image fusion tasks, noise is removed by jointly optimising different modal data through a variational model [298].”

What the sources say

• ?Could not verify

  The 2D-3D VQA Model and Question Answering sections describe fusing 2D images with 3D scene features and jointly optimizing answer predictions, not removing noise from multispectral images. Searches for multispectral fusion, denoising, and variational optimization found no passage addressing the claim; page numbers were not available in our copy. (LLM, conf=0.76)

  cites [298] W. Mo, Y. Liu, Bridging the gap between 2d and 3d visual question answering: A fusion approach for 3d vqa, in: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 4261–4268, 5. sourceExperimental — accuracy not yet measured

In the paper

“using Transformer for text, CNN for image, RNN for audio, etc., and then performs feature alignment through Cross-modal Attention [306,307].”

What the sources say

• ?Not checked — no full text of the sourcecites [306] James, 2020, Rlbench: The robot learning benchmark & learning environment, IEEE Robot. Autom. Lett. source
• Partly supported

  Section B, “3D-LOTUS policy” (page not available), describes a CLIP text encoder and cross-attention that lets point-cloud features attend to word embeddings. It does not describe the claimed combination of a Transformer text encoder, CNN image encoder, and RNN audio encoder; its visual representation is a point cloud encoded by a point-cloud transformer. (LLM, conf=0.88)

  cites [307] Garcia, 2024 sourceExperimental — accuracy not yet measured

In the paper

“Solutions fall into imputation-based [302] and imputation-free methods [303].Imputation-based methods recover missing data via matrix decomposition, graph learning, or GANs, while imputation-free approaches leverage available modalities through latent representation learning and deep learning.”

What the sources say

• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [302] H. Huang, Y. Chen, Z. Wang, R. Huang, R. Xu, T. Wang, L. Liu, X. Cheng, Y. Zhao, J. Pang, et al., Chat-scene: Bridging 3d scene and large language models with object identifiers, in: The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. sourceExperimental — accuracy not yet measured
• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [303] Hong, 2023, 3d-llm: Injecting the 3d world into large language models, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured

In the paper

“For example, encoder-decoder architectures and Transformer-based structures are widely used to model dependencies between modalities [12,13];”

What the sources say

• ?Not checked — no full text of the sourcecites [12] J. Xu, Z. Xiong, S.P. Bhattacharyya, PIDNet: A Real-Time Semantic Segmentation Network Inspired by PID Controllers, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2023, pp. 19529–19539. source
• ?Not checked — no full text of the sourcecites [13] X. Bai, Z. Hu, X. Zhu, Q. Huang, Y. Chen, H. Fu, C.-L. Tai, Transfusion: Robust lidar-camera fusion for 3d object detection with transformers, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1090–1099. source

In the paper

“The integration of diverse modalities enables robots to achieve more robust scene understanding, stronger task generalization, and more natural human-robot interaction in complex environments.”

In the paper

“At the same time, foundation modeldriven multimodal pretraining paradigms are gradually being integrated into robot vision systems, showing great potential for scalability and generalization.”

In the paper

“Traditional multimodal fusion methods offer a basic theoretical framework but struggle with complex data.”

In the paper

“For example, Li et al. [29] presented an evolutionary overview of mainstream VLM architectures, but did not explore their use in tasks like navigation or robotic manipulation.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [29] Li, 2025

In the paper

“With deep neural networks, feature extraction, modality interaction, and decision-making have become deeply integrated, making fusion stages less distinct.”

In the paper

“This has driven a shift from explicit to implicit fusion [37], where network design inherently captures modality relationships, enhancing deep semantic extraction.”

What the sources say

• ?Not checked — no full text of the sourcecites [37] Geiger, 2012, Are we ready for autonomous driving? The KITTI vision benchmark suite source

In the paper

“Occupancy Prediction is vital for multimodal scene understanding, predicting semantic labels for occupied spatial regions, which is crucial for autonomous navigation, obstacle avoidance, and decision-making.”

In the paper

“However, with the advancement of research, more recent approaches have integrated various attention mechanisms with the encoder-decoder framework to enhance the flexibility and accuracy of information fusion.”

In the paper

“However, many occupancy prediction methods rely mainly on visual data, leading to performance drops in adverse conditions like rain or fog.”

In the paper

“Early multimodal fusion methods primarily relied on the encoder-decoder framework to extract and combine information from different modalities.”

In the paper

“In multimodal fusion, the encoder-decoder framework first extracts features from different modalities (e.g., image, text, etc.), with the encoder independently processing each modality to generate high-dimensional representations.The decoder then fuses these features to produce the final output.”

In the paper

“Its parallel computing capability also accelerates the processing of large-scale datasets, making it particularly suitable for real-time scene understanding tasks.”

In the paper

“The introductory survey by Bordes et al. [30] covers training paradigms and evaluation methods, but is mainly intended as a beginnerfriendly guide and fails to address the core challenges of multimodal fusion in robotic vision.”

What the sources say

• ?Not checked — no full text of the sourcecites [30] Wu, 2024, Embodied navigation with multi-modal information: A survey from tasks to methodology, Inf. Fusion source

In the paper

“These attention mechanisms effectively guide the model to focus on key features across modalities, thereby improving the performance of multimodal learning, especially in handling complex tasks.”

In the paper

“By utilizing a unified model architecture, the attention mechanism reduces the complexity of data preprocessing and strengthens the fusion of multimodal data, becoming the mainstream method in the field of multimodal fusion.”

In the paper

“MRFTrans [4], a transformer-based framework, fuses semantic, geometric, and depth information to address the challenge of monocular 3D scene completion, achieving state-of-the-art performance on SemanticKITTI and NYUv2 datasets by effectively modeling long-range dependencies and refining depth estimates.”

What the sources say

• ?Not checked — no full text of the sourcecites [4] Xu, 2024, MRFTrans: Multimodal representation fusion transformer for monocular 3D semantic scene completion, Inf. Fusion source

In the paper

“Post fusion is a key method in multimodal analysis, which combines the results of decision level indepen-dent processing of modalities [45].”

What the sources say

• ?Not checked — no full text of the sourcecites [45] P. Li, X. Li, Multimodal Fusion with Co-attention Mechanism, in: 2020 IEEE 23rd International Conference on Information Fusion, FUSION, 2020, pp. 1–8. source

In the paper

“Early fusion directly integrates multimodal inputs at the data level, such as concatenating RGB [40] images and depth maps before feeding them into the model.”

What the sources say

• ?Not checked — no full text of the sourcecites [40] Zhao, 2024, Deep multimodal data fusion, ACM Comput. Surv. source

In the paper

“During this process, cross-modal feature fusion is typically achieved through early or deep fusion, often optimized by incorporating attention mechanisms to enhance the flow of information and focus on key features.”

In the paper

“These approaches collectively advance multimodal fusion for semantic scene understanding.”

In the paper

“Common techniques include weighted averaging, voting mechanisms, and logical rules .”

In the paper

“Methods like OccuFusion [62] use SENet modules [63] to adaptively weight modality features, improving generalization and reliability.”

What the sources say

• ?Not checked — no full text of the sourcecites [62] Daneshfar, 2025, Elastic deep multi-view autoencoder with diversity embedding, Inform. Sci. source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [63] Li, 2024

In the paper

“Its advantage lies in its simple structure and the ability to directly utilize the raw data characteristics, while its disadvantage is that the features between modalities may not match, making it difficult to handle high-dimensional data.”

In the paper

“Mid term fusion occurring in the middle layer is a common method for balancing modal independence and capturing semantic relationships.”

In the paper

“GNN-based methods [52,53] model multimodal data through graph structures, enhancing global and robust semantic understanding.”

What the sources say

• ?Not checked — no full text of the sourcecites [52] Liu, 2022, Focus your attention: A focal attention for multimodal learning, IEEE Trans. Multimed. source
• ?Not checked — no full text of the sourcecites [53] R. Kuga, A. Kanezaki, M. Samejima, Y. Sugano, Y. Matsushita, Multi-task Learning Using Multi-modal Encoder-Decoder Networks with Shared Skip Connections, in: 2017 IEEE International Conference on Computer Vision Workshops, ICCVW, 2017, pp. 403–411. source

In the paper

“This method has multiple advantages, such as strong modal independence, ease of individual optimization, and scalability of multimodal systems [46] .”

What the sources say

• ?Not checked — no full text of the sourcecites [46] Zhang, 2024, Multimodal remote sensing image matching via learning features and attention mechanism, IEEE Trans. Geosci. Remote Sens. source

In the paper

“This fusion reduces sensor limitations (e.g., occlusions, noise) and enhances scene interpretation robustness.”